Channel types predictions for the Sacramento River basin

Hydrologic and geomorphic classifications have gained traction in response to the increasing need for basin-wide water resources management. Regardless of the selected classification scheme, an open scientific challenge is how to extend information from limited field sites to classify tens of thousands to millions of channel reaches across a basin. To address this spatial scaling challenge, we leveraged machine learning to predict reach-scale geomorphic channel types using publicly available geospatial data.

Methods
A bottom-up machine learning approach selects the most accurate and stable model among ~96,000 and derives the relationship between 147 predictors and labels corresponding to regional channel types in a three-tiered framework which: (i) define a tractable problem; assess model performance (ii) in statistical learning; and (iii) in prediction. In the present application to the Sacramento River basin (California, USA), the developed framework selects a Random Forest model to predict 10 channel types previously determined from 290 field-surveys over 108,943 200-m stream interval.

Performance in statistical learning is high with a 65% median cross-validation accuracy and a 0.91 mean multiclass Area Under Curve value. Furthermore, the predictions coherently capture the expected geomorphic organization of the landscape. As main metric of uncertainty, we include for each stream-segment the entropy calculated from the posterior probabilities output from the machine learning algorithm. For completeness, evenness and richness are also reported.

The predictions included in this dataset corresponds to an aggregated version of the output from the machine learning framework. Each initial 200-m stream interval was aggregated by using their COMIDs from the National Hydrography Dataset corresponding to the most common identifier used by stake-holders. For each resulting NHD stream line, the 200-m scale probabilities associated to each channel types are summed and normalized to sum to one.